[dpdk-dev] VMWare Performance - vmxnet3-usermap

[Stefan Baranoff]

All,

I've been playing with DPDK recently on a variety of bare metal Linux
installations and so far and have seen wonderful improvements in
performance on both our Westmere and Sandy Bridge based servers.

However when I install ESXi 5.1 (not linked to a vSphere management system
-- stand alone ESXi installation) on one of the Westmere systems and use
vmxnet3-usermap with the standard VMWare vSwitch my performance drops way
down.

Does anyone have a sense of pps/bps I can realistically expect to see from
vmxnet3-usermap without doing SR-IOV/passthrough? Raw CentOS and RHEL 6.4
we're seeing 14.88Mpps/10Gbps but going to ESXi running CentOS 6.4 we're
seeing 500Kpps/4Gbps. Is that reasonable (obviously packet rate is with
small packets and data rate is with larger packets). Without going to
SR-IOV is there anything that I can do to improve this performance in
vmware?

Also, I know SR-IOV breaks many of the HA/auto balancing features of
VMware. Is the same true with vmxnet3-usermap or is that safe to use with
VMs floating around a cluster willy-nilly?


Thanks,
Stefan